Skip to content

TEST.md

Latest commit

 

History

History
86 lines (70 loc) · 3.24 KB

TEST.md

File metadata and controls

86 lines (70 loc) · 3.24 KB

some test results from https://github.com/BobLiu20/Classification_Nets

Nets

Kind in mind, it is not any augmentation in training.
The image size of cifar10 is 32x32.

caffe: 1.0.0-rc3
cuda: 8.0
nvidia: GTX1080 ti (Total memory is 11GB)
system: ubuntu 14.04 in docker

  • LeNet
    The size of model: 351KB
    The accuracy of testing is 0.7823 after 64000 Iterations.
    The time of training is 0.5 hours in GTX1080.
    The time of predict is 0.93 ms. (one image with 32x32)

  • BN-LeNet: Batch Normalization LeNet
    The size of model: 352KB
    The accuracy of testing is 0.7935 after 64000 Iterations.
    The time of predict is 0.76 ms. (one image with 32x32)

  • AlexNet
    The size of model: 3MB
    The accuracy of testing is 0.7452 after 64000 Iterations.
    The time of training is 10 mins in GTX1080.
    The time of predict is 0.65 ms. (one image with 32x32)

  • SqeezeNet_v1.1
    The size of model: 2.8M
    The accuracy of testing is 0.8114 after 64000 Iterations.
    The time of predict is 2.2 ms. (one image with 32x32)

  • NetworkInNetwork: NIN
    The size of model: 25MB
    The accuracy of testing is 0.8346 after 64000 Iterations.
    The time of training is 40 mins in GTX1080.
    The time of predict is 1.8 ms. (one image with 32x32)

  • ResNet20
    The size of model: 1.1MB
    The accuracy of testing is 0.8258 after 64000 Iterations.
    The time of predict is 4.2 ms. (one image with 32x32)

  • ResNet32
    The size of model: 1.8MB
    The accuracy of testing is 0.8794 after 64000 Iterations.
    The time of predict is 7.8 ms. (one image with 32x32)

  • ResNet56
    The size of model: 3.4MB.
    The accuracy of testing is 0.8706 after 64000 Iterations.
    The memory usage of GPU is 3.8GB in training. (batch size is 128).
    The time of training is 15 hours in GTX1080.
    The time of predict is 16.3 ms. (one image with 32x32)

  • WRN28_10: Wide Residual Networks
    The size of model: 140MB.
    The accuracy of testing is 0.8950 after 60000 Iterations.
    The memory usage of GPU is 9.9GB in training. (batch size is 128).
    The time of training is 22.5 hours in GTX1080 ti.
    The time of predict is 13.0 ms. (one image with 32x32)

  • VGG16
    The size of model: 129MB.
    The accuracy of testing is 0.8308 after 64000 Iterations.
    The memory usage of GPU is 1GB in training. (batch size is 128).
    The time of training is 1 hours in GTX1080 ti.
    The time of predict is 4.6 ms. (one image with 32x32)

  • GoogLeNet
    The size of model: 25MB.
    The accuracy of testing is 0.7913 after 64000 Iterations.
    The memory usage of GPU is 1.3GB in training. (batch size is 128).
    The time of training is 1 hours in GTX1080 ti.
    The time of predict is 7.2 ms. (one image with 32x32)

  • DenseNet: (Ref to here)
    The size of model: 4MB.
    The accuracy of testing is 0.9153 after 64000 Iterations.
    The memory usage of GPU is 7.9GB in training. (batch size is 32+32).
    The time of training is 3.5 hours in GTX1080 ti.
    The time of predict is 12.95 ms. (one image with 32x32)